A robot that can pour coffee is easy to film. A robot that can pour coffee reliably, cheaply, and at scale is a different country's industrial problem. Beijing just attached money and a deadline to closing that gap.
The document, in plain terms
On February 28, 2025, three Beijing bodies — the Municipal Science and Technology Commission (also Zhongguancun Administrative Committee), the Development and Reform Commission, and the Economy and Information Technology Bureau — issued the Beijing Embodied Intelligence (具身智能) Science, Technology Innovation and Industry Cultivation Action Plan (2025–2027) (京科信发〔2025〕36号). It runs through December 31, 2027, and reads less like a vision statement and more like a construction schedule.
Embodied intelligence (具身智能) here means AI that acts through a physical body — robots that perceive, decide, and move, rather than models that only answer text. Beijing is betting that the next platform shift after software is the machine that shares our spaces.
The 2027 scorecard
The plan sets a specific set of numbers for 2027:
- Break through no fewer than 100 key technologies, concentrated on embodied brain-and-cerebellum models, embodied-intelligence chips, and whole-body motion control.
- Produce no fewer than 10 internationally leading software or hardware products.
- Make the upstream and downstream supply chain basically localized (国产化).
- Build a batch of new research platforms — world-model simulation, data collection, pilot verification, and open scenario testing — and a unified data-collection-management and test-verification standard for embodied data.
- Support no fewer than 100 innovation entities and no fewer than 50 core enterprises across the chain.
- Yield no fewer than 50 mass-production products and no fewer than 100 scaled applications across three scenarios: research and education, industry and commerce, and personalized services.
- Push cumulative mass production past 10,000 units (万台) and grow a hundred-billion-yuan (≈ US$14B / HK$110B) industrial cluster.
- Build no fewer than 2 specialized industrial clusters and a industry-education fusion base.
Where the money is
The plan is backed by real budget, not just prose. Through municipal fixed-asset investment, Beijing will support embodied-intelligence technology breakthroughs, data work, and application demonstration, with a single project eligible for up to ¥100M (≈ US$14.1M / HK$110M). Two existing vehicles — the municipal AI-industry investment fund and the robotics-industry development investment fund — are pointed at early hard-tech in the sector.
That per-project ceiling matters: it signals Beijing would rather subsidize a few deep, capital-intensive builds than scatter small grants.
Three scenarios, three pressures
The "100 scaled applications" target is split across research and education, industry and commerce, and personalized services. Each carries a different bar. A university lab demo is forgiving. A warehouse deployment is not. A home or care robot that fails in front of a person is the unforgiving one. The plan names all three because the hard part of embodied intelligence (具身智能) is not the model — it is the last meter of real-world reliability.
Standards before scale
One line easy to miss: Beijing wants a unified data-collection-management standard and a test-verification standard for embodied data. Robots learn from messy real-world interaction; without a common data format and a common way to test a behavior, every company rebuilds the same scaffold. The plan treats standardization as infrastructure, not paperwork.
Why a city, not just the nation
National industrial policy sets direction; cities compete to host the build-out. Beijing's angle is its research density — the universities, the labs, the prototype shops. The plan is essentially a retention and attraction strategy: keep the breakthroughs local, and make the cluster sticky with funding, space, and testbeds.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the 10,000-unit target is less a market forecast than a forcing function — a number big enough to expose which suppliers and which data pipelines are actually ready.
Honest limitations
These are the issuer's own targets, not independent projections, and "basically localized (国产化)" is a qualitative phrase with no published threshold. The 10,000-unit figure is cumulative mass-production scale, and how it is counted (pilot batches vs. commercial shipments) will shape whether it looks met. The ¥100M per-project ceiling depends on implementing rules and annual budget that can shift. We have not verified a 2025 baseline report, so progress toward these numbers cannot yet be confirmed from public sources.
What readers can do now
- For startups: the plan names open "list-and-race" (揭榜挂帅) mechanisms and incubators — relevant if you build actuators, embodied chips, or data tooling rather than whole robots.
- For investors: the two named Beijing funds are the clearest signal of where local capital will lean; co-investment windows tend to open right after plan issuance.
- For researchers: the world-model simulation and data-collection platforms are explicitly meant to serve outside innovators — a reason to watch for open calls.
- For observers: check back near the 2027 deadline for an official progress accounting before judging whether the hundred-billion-yuan (≈ US$14B / HK$110B) cluster materialized or merely got announced.